Ensemble graph auto-encoders for clustering and link prediction.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph auto-encoders generate node embeddings for unsupervised learning.
• Existing models struggle with contextual node information, leading to poor embeddings.
• There is a need for improved graph representation learning methods.

Purpose of the Study:

• To propose the Ensemble Graph Auto-encoders (E-GAE) model.
• To enhance the quality of node embeddings in graph data.
• To improve performance on downstream graph learning tasks.

Main Methods:

• The E-GAE model integrates three techniques: ensemble random walk graph auto-encoder, random walk graph auto-encoder of the ensemble network, and graph attention auto-encoder.
• Node embedding matrices are generated and combined using adaptive weights.
• The model reconstructs a new node embedding matrix to mitigate embedding quality issues.

Main Results:

• Experiments on Cora, Citeseer, and PubMed datasets demonstrate E-GAE's effectiveness.
• The model achieved up to a 2.0% improvement in link prediction.
• A 9.4% enhancement was observed in clustering tasks.

Conclusions:

• The proposed E-GAE model effectively addresses limitations in existing graph auto-encoders.
• E-GAE generates higher-quality node embeddings by capturing richer contextual information.
• The method shows significant performance gains in link prediction and clustering.